Rise of the Machines: Algorithmic Trading in the Foreign Exchange Market Alain Chaboud Benjamin Chiquoine Erik Hjalmarsson Clara Vega [PRELIMINARY AND INCOMPLETE] February 5, 29 Abstract We study the impact that algorithmic trading, computers directly interfacing with trading platforms, has had on price discovery and volatility in the foreign exchange market, using high frequency data representing a majority of global interdealer trading in ve major currency pairs from 26 to 27. Our dataset contains precise observations of the fraction and the direction of the computer-generated trades each minute. As such, it allows us to analyze the possible links between algorithmic trading and market volatility, to identify whose trades have a more permanent impact on prices, and to study how correlated algorithmic trades are. We nd that non-algorithmic order ow accounts for most of the (long-run) variance in exchange rate returns, i.e. non-algorithmic traders are better informed. We also nd that there is, in some cases, an over-reaction of the price to algorithmic order ow. There is some evidence that algorithmic trades tend to be correlated, suggesting that the algorithmic strategies used in the market may not be as diverse as those used by non-algorithmic traders. JEL Classi cation: F3, G12, G14, G15. Keywords: Algorithmic trading, excess volatility, private information. The authors are a liated with the Division of International Finance at the Federal Reserve Board of Governors. Please address comments to the authors via e-mail at alain.p.chaboud@frb.gov, bchiquoine@gmail.com, erik.hjalmarsson@frb.gov and clara.vega@frb.gov. The views in this paper are solely the responsibility of the authors and should not be interpreted as re ecting the views of the Board of Governors of the Federal Reserve System or of any other person associated with the Federal Reserve System.
1 Introduction The use of algorithmic trading, where computer algorithms directly manage the trading process at high frequency, has become common in major nancial markets in recent years, beginning in the U.S. equity market more than 15 years ago. There has been widespread interest in understanding the potential impact of algorithmic trading on market dynamics, as some analysts have highlighted the potential for improved liquidity and more e cient price discovery, while others have expressed concern that it may be a source of increased volatility and reduced liquidity, particularly in times of market stress. 1 Despite this interest, however, there has been almost no formal empirical research on the topic, primarily because of a lack of data where algorithmic trades are clearly identi ed. A notable exception is a recent paper by Hendershott, Jones, and Menkveld (27), who get around the data constraint by using the ow of electronic messages on the NYSE as a proxy for algorithmic trading. They conclude that algorithmic trading on the NYSE, contrary to the pessimists concerns, likely causes an improvement in market liquidity. 2 In the foreign exchange market, the adoption of algorithmic trading (AT) is a far more recent phenomenon than in the equity market, as the two major interdealer electronic trading platforms only began to allow algorithmic trades a few years ago. Growth in AT has been rapid, however, and a sizable fraction of foreign exchange transactions currently involve at least one algorithmic counterparty. We study in this paper the impact that AT has had on price discovery and volatility using high-frequency data representing a majority of global interdealer trading in ve major currency pairs from January 26 to December 27, a period over which the share of AT in the foreign exchange markets rose rapidly. Importantly, our dataset contains precise observations of the fraction and the direction of the computer-generated trades each minute. As such, it allows us to study how algorithmic trading and market volatility are related, to identify whether algorithmic or non-algorithmic trades have a more permanent impact on prices, and to estimate how correlated algorithmic trades are. In algorithmic trading, computers directly interface with trading platforms, placing orders without human intervention. The computers observe market data and possibly other information at very high frequency, and, based on a built-in algorithm, send back trading instructions. A variety of algorithms are used: some look for arbitrage opportunities, for instance small discrepancies in the exchange rates between three currencies; some seek optimal execution of large orders at the minimum cost; and some seek to implement longer-term 1 For instance, an article published by the Financial Times on December 5, 28, was titled "Algorithmic trades produce snowball e ects on volatility." 2 We also note a paper by Hasbrouck (1996) on program trading, where he analyzes 3 months of data where program trades can be separately identi ed from other trades. He concludes that both types of orders have an approximately equivalent impact on prices. Algorithmic trading is not exactly equivalent to program trading, though it is a close cousin. In principle, a program trade could be generated by a trader s computer and then the trade conducted manually by a human trader. Our de nition of AT refers to the direct interaction of a trader s computer with an electronic trading platform, that is the automated placement of a trade order on the platform. 1
trading strategies in search of pro ts. Among the most recent developments in algorithmic trading, some algorithms now automatically read and interpret economic data releases, generating trading orders before economists have nished reading the rst line. 3 The extreme speed of execution that AT allows and the potential that algorithmic trades may be highly correlated, perhaps as many institutions use similar algorithms, have been cited as reasons for concerns, as some have feared that AT may generate large price swings and market instability. One such instance may have happened on August 16, 27, in a period of extreme volatility, the highest in our sample period. On that day, the Japanese yen appreciated sharply against the U.S. dollar around 6: a.m. and 12: p.m. (NY time) as we show in Figure 1. The gure also shows, for each 3-minute interval in the day, algorithmic ("computer") order ow in the top panel and non-algorithmic ("human") order ow in the lower panel. The two sharp exchange rate movements mentioned happened when computers, as a group, aggressively sold dollars and bought yen. Human order ow at those times was, in contrast, small. Humans traders then aggressively bought dollars after 12: p.m, and the appreciation of the yen against the dollar was partially reversed. This is only a single example, of course, but it leads us to ask whether computer trades tend to create excess volatility, in the sense that exchange rate movements driven by computer trades are more likely to be later reversed. This example also leads us to ask whether human trades routinely have a more permanent impact on prices than computer trades. We formally investigate these conjectures using minute-by-minute data from January 26 to December 27 on ve exchange rate pairs: the euro-dollar, dollar-yen, dollar-swiss franc, euro-yen, and euro-swiss franc. We nd that, controlling for potential endogeneity biases and for the common trend in exchange rate volatility and algorithmic trading, there is no evident causal relationship between AT and volatility. However, the instruments we use in the analysis are weak and thus we also analyze return-order ow dynamics in a high-frequency VAR framework in the tradition of Hasbrouck (1991a). The VAR estimation provides three important insights. First, we nd that human order ow accounts for most of the (long-run) variance in exchange rate returns, i.e., humans are the informed traders in these markets. This may partially be attributed to the fact that some of the algorithmic trading is used for the optimal execution of large orders at a minimum cost. Algorithmic trades appear to be successful in that endeavor, with computers breaking up the larger orders and having a minimum impact on prices. Second, we nd that, on average, computers or humans that trade on a price posted by a computer do not impact prices quite as much as they do when they trade on a price posted by a human. One possible interpretation of this result is that this is evidence that computers tend to place limit orders more strategically than humans do. This nding may relate to the literature that proposes to depart from the 3 The Economist, June 21, 27 2
prevalent assumption that liquidity providers in limit order books are passive. 4 Third, our VAR analysis shows that there is an initial under-reaction to order ow between humans (where the price is both posted by and dealt on by a human), while there is an initial over-reaction to order ow between computers. The euro-dollar exchange-rate pair during our three-month subsample provides an extreme example, as the initial reaction to computer-computer order ow is a 21 basis point move, but the long-run cumulative reaction is just 4 basis points. To the extent that there is an initial over-reaction to computer-computer order ow, we conclude that algorithmic trading may be linked to some excess short-run volatility. Also, in the euro-dollar and dollar-yen markets, the presence of a computer as a liquidity provider or a liquidity demander is linked to some short-term overreaction of the price. But this is not the case for the dollar-swiss, the euro-yen, and the euro-swiss franc exchange rates. Coincidentally, these are the exchange rates where, by the end of our sample, AT is as prevalent or more prevalent than human trading. We believe that a substantial fraction of the AT in these markets re ects computers taking advantage of so-called triangular arbitrage opportunities, where the prices set in, say, the euro-dollar and dollar-yen markets are very brie y out of line with the eur-yen rates. In these cases, computer trading likely contributes to market e ciency by narrowing misalignments of exchange rates. Finally, we nd some evidence that, in all our currency pairs, computer trades are more highly correlated with each other than human trades, suggesting that the strategies used by computers are not as diverse as those used by humans. This fact echoes the concerns voiced by some analysts that, as computers take over trading in nancial markets, these markets will miss the bene ts of the divergence of opinion among humans, as well as their slower reaction times and perhaps more subtle judgment. However, since the high correlation of computer trades does not seem to automatically translate into economically signi cant excess volatility, it is not clear how damaging that high correlation is. We proceed as follows. In Section 2 we describe the Electronic Broking Services (EBS) exchange rate data, describing the evolution over time of algorithmic trading and the pattern of interaction between human and algorithmic traders. In Section 3 we study the relationship between realized volatility and the activity of algorithmic traders. The econometric techniques used in this section take advantage of di erences in the time series of volatility and AT prevalence among the di erent currency pairs to address likely endogeneity issues. In Section 4 we analyze return-order ow dynamics in a VAR framework to identify whose trades, computers or humans, have a more permanent impact on prices. In Section 5 we examine how correlated the algorithmic orders are with each other compared to the human orders. In Section 6 we conclude. 4 For example, Chakravarty and Holden (1995), Kumar and Seppi (1994), Kaniel and Liu (26), and Goettler, Parlour and Rajan (27) allow informed investors to use both limit and market orders. Bloom eld, O Hara and Saar (25) argue that informed traders are natural liquidity providers and Angel (1994) and Harris (1998) show that informed investors can optimally use limit orders when private information is su ciently persistent. 3
2 Data description Today, two electronic platforms process the vast majority of global interdealer spot trading in all the major currency pairs, one o ered by Reuters, and one o ered by EBS. These platforms, which are both electronic limit order books, have become essential utilities for the foreign exchange market. Importantly, trading in each major currency pair has become over time very highly concentrated on only one of the two systems. Of the most traded currency pairs, the top two, euro-dollar and dollar-yen, trade primarily on EBS, while the third, sterling-dollar trades primarily on Reuters. As a result, the reference price at any moment for, for example, spot euro-dollar is the current price on the EBS system, and all dealers across the globe base their customer and derivative quotes on that price. We have access to AT data from EBS from 23 through 27. We focus, however, on the sample from 26 to 27, because, as we show in Figure 2, algorithmic trades were a very small portion of all trades in the earlier years. In addition to the full 26-27 sample, we also consider a sub-sample covering the months of September, October, and November of 27. Since the growth in algorithmic trading continued throughout 26 and 27, it is interesting to separately analyze these three months towards the end of the sample period, when algorithmic trading played an even more important role than earlier in the sample. 5 We have access to the ve most-traded currency pairs on the EBS system: euro-dollar, dollar-yen, euro-yen, dollar-swiss franc, and euro-swiss franc. The quote data, at the one-second frequency, consist of the highest bid quote and the lowest ask quote on the EBS system in these currency pairs, from which we construct mid-quote series and compute one-minute exchange rate returns. The transactions data, at the one-minute frequency, consist, for each currency pair, of the amounts of base currency bought and sold. We can also identify the type of trader, human or computer, who posted the price at which the transaction was conducted (the maker ) and the type of trader who decided to buy or sell at that price (the taker ). 6 The main goal of this paper is to analyze the e ect algorithmic trading has on price discovery and volatility. To that end, we analyze di erent decompositions of "order ow" (we clearly stretch the traditional de nition of order ow, as shown below). First, we decompose order ow into the four most disaggregate components: human-maker/human-taker (HH), computer-maker/human-taker (CH), human-maker/computer-taker (HC), and computer-maker/computer-taker (CC). Second, we decompose order ow into the standard separation, which distinguishes trades based on who initiated the trade: human-taker (HH+CH) and computer-taker (HC+CC). Third, we decompose order ow into a separation that distinguishes trades based on who provides liquidity: human-maker (HH+HC) and computer-maker (CH+CC). Fourth, we decompose order ow into 5 We do not use December 27 in the sub-sample to avoid the in uence of year-end e ects. 6 There is a very high correlation in this market between trading volume per unit of time and the number of transactions per unit of time, and the ratio between the two does not vary much over time. Order ow measures based on amounts transacted and those based on number of trades are therefore very similar. 4
purely human trades (HH) and trades where at least one of the two counterparties was an algorithmic trader (CH+HC+CC). The rst decomposition allows us to analyze the e ect order ow has on prices when, for instance, no party has a speed advantage, i.e. both parties are humans or both parties are computers, and when either the maker has a speed advantage, CH, or the taker has a speed advantage, HC. This distinction may be particularly useful when analyzing the cross-rates, where computers likely have a clear advantage over humans in detecting short-lived triangular arbitrage opportunities. This decomposition may also allow us to study whether the liquidity supplier, who is traditionally assumed to be uninformed, is posting quotes strategically. This situation is more likely to arise in our database, a pure limit order book market, than in a hybrid market like the NYSE, because, as Parlour and Seppi (28) point out, the distinction between liquidity supply and liquidity demand in limit order books is blurry. 7 Still, in our exchange rate data as in other nancial data, the net of trades signed by who the taker is (the standard de nition of order ow) is clearly highly positively correlated with exchange rate returns, so that the taker is considered to be more "informed" than the maker. Thus we also consider prominently the more traditional second decomposition, human-taker and computer-taker order ow, in our analysis of whose trades impact prices more. The third decomposition, human-maker and computer-maker order ow, allows us, for instance, to determine whether computers or humans are more likely to provide liquidity when it is needed the most, e.g. during periods of high exchange rate volatility. Lastly, the fourth decomposition, computer-participation and human-human order ow, allows us to determine whether any type of participation by computers, passive or active, is linked to excess volatility in the market. In our analysis, we exclude data collected from Friday 17: through Sunday 17: New York time from our sample, as activity on the system during these non-standard hours is minimal and not encouraged by the foreign exchange community. We also drop certain holidays and days of unusually light volume: December 24-December 26, December 31-January 2, Good Friday, Easter Monday, Memorial Day, Labor Day, Thanksgiving and the following day, and July 4 (or, if this is on a weekend, the day on which the Independence Day holiday is observed). We show summary statistics for the one-minute returns and order ow data in Table 1 and Table 2. These tables contain a number of noteworthy features. First, order ow is serially positively correlated. This is consistent with informed trading models. For example, Easley and O Hara (1987) model a situation where sequences of large purchases (sells) arise when insiders with positive (negative) signals are present in the market. He and Wang (1995) also show that insiders with good (bad) news tend to buy (sell) repeatedly 7 Parlour and Seppi (28) note that in a limit order book investors with active trading motives, some of which are informed traders, may choose to post limit orders that are more aggresive than those a disinterested liquidity provider would use but less aggresive than market orders. 5
until their private information is revealed in the prices. The positive serial correlation in order ow is also consistent with strategic order splitting, i.e. a trader willing to buy for informational or noninformational reasons and splitting his order to reduce market impact. The serial correlation in order ow is of additional interest to us because we will need to control for it in our regression speci cations. Second, the standard deviation of order ow is di erent across maker/taker pairs and exchange rates. For example, in the two exchange rate markets with the highest trading volume, the euro-dollar and dollar-yen markets, the standard deviation of human-taker order ow is larger than the standard deviation of computer-taker order ow. A consequence is that, in these two markets, large one-sided market orders are more likely to be executed by human takers than by computer takers. Di erences in standard deviations across maker/taker order ow pairs are important to the interpretation of the VAR analysis. It could be the case, for instance, that the price impact of a one billion dollar shock to CC order ow is the same as the price impact of a one billion dollar shock to HH order ow. However the percent of the total variation in exchange rate prices explained by the latter type of order ow would be larger because its standard deviation is larger. The correlations between the most disaggregate types of order ow are shown in Table 3, both for the full 26-27 sample as well as for the shorter three month sub-sample. One notable result in these tables is that the four types of order ow are not highly correlated (positively or negatively) except for the HC and CH order ows, with a correlation of about -.4. This is consistent with Parlour and Seppi (28) s assertion that, in a limit order book, investors with active trading motives may choose to place limit orders that are more agressive than those a disinterested liquidity provider would place. In other words, when computers, for example, want to buy dollars and sell euros they will not only do it by executing market orders but they will also post limit orders that are aggressive and are more likely to be picked up by humans than by other computers, i.e. when HC is positive CH tends to be negative. 8 We show in Figure 2, from 23 through 27, for the ve major currency pairs trading on EBS, the fraction of trades where at least one of the two counterparties was an algorithmic trader (CH+HC+CC). From its beginning in the second half of 23, the fraction of trades involving AT grew by the end of 27 to near 6% for euro-dollar, dollar-yen, and euro-swiss trading, and to about 8% for euro-yen and dollar-swiss. Figure 3 shows, for each of our ve currency pairs, the evolution over time of the four di erent possible types of trades. By the end of 27, in the euro-dollar market, human to human trades, in black, accounted for slightly less than half of the volume, and computer to computer trades, in green, for about ten percent. Computers took prices posted by humans about as often as humans took prices posted by market-making computers, in blue. The same pattern is also found in the dollar-yen market. Since the presence of more 8 We also note that, since the correlation across di erent types of order ow is not extremely high, with perhaps the exception of CH and HC, we can have less concern about multicollinearity in some of our regression speci cations. 6
makers increases market liquidity, i.e., larger trades can be executed with little impact on the price, Figure 3 shows that in the most-traded currency pairs, computer and human traders contributed about evenly to market liquidity. The story is di erent for the cross-rate, the euro-yen currency pair. By the end of 27, there were more computer to computer trades than human to human trades, and the most common type of trade was computers trading on prices posted by humans (HC). We believe this re ects computers taking advantage of triangular arbitrage opportunities, where prices set in the euro-dollar and dollar-yen markets are very brie y out of line with the euro-yen cross rate. Trading volume is largest in euro-dollar and dollar-yen markets, and price discovery happens mostly in those markets, not in the cross-rate. The dollar-swiss franc and euro-swiss franc markets are also more highly reliant on AT by the end of 27. 9 3 The impact of algorithmic trading on volatility In this section we attempt to estimate whether the presence of algorithmic trading causes disruptive market behavior in the form of increased volatility. 3.1 Identi cation The main challenge in identifying a causal relationship between algorithmic trading and volatility is the potential endogeneity of algorithmic trading with regards to variables such as volatility. That is, although one may conjecture that algorithmic trading impacts volatility, it is also highly plausible that algorithmic trading activity is a function of the level of volatility. For instance, highly volatile markets may present comparative advantages to automated trading algorithms relative to human traders, which might increase the fraction of algorithmic trading during volatile periods. In contrast, however, one could also argue that a high level of volatility might reduce the informativeness of historical price patterns on which some trading algorithms are likely to base their decisions, and thus reduce the e ectiveness of the algorithms and lead them to trade less. The bottom line is that the fraction of algorithmic trading is likely to be endogenous with regards to exchange rate volatility. We cannot easily determine in what direction the bias will go in an OLS regression of volatility on the fraction of algorithmic trading, because it is not obvious whether higher volatility would induce more or less algorithmic trading. To deal with the endogeneity issue, we adopt an instrumental variable (IV) approach as outlined below. 9 The foreign exchange markets for the Swiss franc are highly dependent on the trading activity of the two large Swiss banks, UBS and Credit Suisse, which are known for their sophisticated electronic trading activity. Traders tell us that, in the Swiss franc exchange markets, the dollar-swiss franc pair is generally viewed as the "third leg" of the triangular arbitrage play, with price discovery occuring primarily in euro-dollar and euro-swiss franc. 7
We are interested in estimating the following regression equation, RV i;t = i + i AT i;t + ix i;t + i;t ; (1) where i = 1; :::; 5 represents currency pairs and t = 1; :::; T, represents time. RV i;t is (log) realized daily volatility, AT i;t is the fraction of algorithmic trading at time t in currency pair i, x i;t is a set of control variables that will primarily contain lagged values of RV i;t as well as time dummies that control for secular trends in the data, and i;t is an error term that is assumed to be uncorrelated with x i;t but not necessarily with AT i;t. The exact de nitions of RV i;t, AT i;t, and x i;t will be given later. The main focus of interest is the parameter i, which measures the impact of algorithmic trading on RV i;t in currency pair i. However, since AT i;t and i;t may be correlated, due to the potential endogeneity discussed above, the OLS estimator of i may be biased. In order to obtain an unbiased estimate, we will therefore consider an instrumental variable approach. Formally, we need to nd a variable, or set of variables, z i;t, that is uncorrelated with i;t (validity of the instrument) and correlated with AT i;t (relevance of the instrument). Thus, z i;t needs to be uncorrelated with the variation left in volatility, after controlling for the variation in x i;t and AT i;t. The starting point of our identi cation scheme is the fact that we have data on several currency pairs. A natural instrument for AT i;t that comes to mind is therefore algorithmic trading in the other currency pairs fat j;t g j6=i. However, since volatility is correlated across currency pairs, it is likely that i;t is also correlated across currency pairs. Thus, under the assumption that AT i;t and i;t are correlated, it follows that it is likely that AT j;t and i;t are also correlated and AT j;t may therefore not be a valid instrument. Instead, we propose to use the lagged values of AT j;t as instruments; that is, fat j;t 1 g j6=i. Since there is both serial correlation and cross-correlation across currencies in the fraction of algorithmic trading, these instruments should be relevant; i.e., correlated with AT i;t. Importantly, however, these lagged variables are also likely to be valid instruments when x i;t is de ned appropriately. For instance, let x i;t include lagged values of both RV i;t and frv j;t g j6=i. The lags of own volatility are used to control for the well known serial correlation in volatility. The lags of the volatility for the other currency pairs are included to ensure the validity of the proposed instruments. That is, by controlling for lagged values of volatility in all currency pairs, the error term i;t should only be contemporaneously correlated with the volatility in other currencies, and not with the lagged values. Consequently, we would also expect i;t to be uncorrelated with the lagged values of algorithmic trading in other currencies, fat j;t 1 g j6=i, which suggests that these should provide valid instruments. Since the cross-currency exchange rates (the euro-franc and the euro-yen) are e ectively determined by the other three main currency pairs in the sample, there might be some concern that this 8
would a ect the validity of the above IV approach. In addition to performing the IV estimation using all ve currency pairs, we therefore also repeat the estimation using only the three main currency pairs. The instrumental variable regressions are estimated using Limited Information Maximum Likelihood (LIML), and we test for both the relevance and the validity of the instruments by reporting the Stock and Yogo (25) test of weak instruments and the standard J test of overidentifying restrictions, which provides a test of the instrument validity. We use LIML rather than two-stage OLS since Stock and Yogo (25) show that the former is much less sensitive to weak instruments than the latter. Another inferential issue, quite distinct from the endogeneity issue just discussed, is the strong upwards trend in the fraction of algorithmic trading over the sample period. As seen in the previous graphs, this trend is clearly the dominant feature of the time-series behavior of the fraction of algorithmic trading. Thus, if one does not attempt to control for it, any regression results with algorithmic trading as an explanatory variable will primarily re ect the correlation between the left-hand-side variable and this increasing secular trend. For instance, if there is a tendency for the left-hand-side variable to trend downwards over the sample period, as is the case for the volatility in some currency pairs, then the estimated slope-coe cient will most likely be negative. However, although one cannot rule out that there is therefore a long-run negative relationship between volatility and algorithmic trading, it is also quite possible that the downward trend in volatility is driven by some other factor that is not accounted for in the model. Since there is no feasible way to control for all other potential factors that may have caused long-term shifts in the level of volatility, we focus on the impact of changes in algorithmic trading from some local mean that changes over time. In particular, monthly time dummies are included as control variables. The regression results should therefore be interpreted as the impact of changes in algorithmic trading over shorter time periods and not the e ect of going from virtually no algorithmic trading to, say, 3 4 percent of the total trading volume. The latter question is arguably as interesting as the former, but extremely di cult to answer without very strong assumptions. 3.2 Variable de nitions 3.2.1 Realized Volatility Volatility is measured as the daily realized volatility obtained from ve minute returns; that is, the volatility measure is equal to the daily sum of squared ve minute log-price changes. The use of realized volatility, based on high-frequency intra-daily returns, as an estimate of ex post volatility is now well established and generally considered the most precise and robust way of measuring volatility. We use ve minute returns to avoid any bias in the estimation of volatility, which may arise from market microstructure noise present in returns sampled at even higher frequencies (e.g. Hansen and Lunde, 26). Following the common conventions in 9
the literature on volatility modelling (e.g. Andersen et al., 21), the realized volatility is log-transformed to obtain a more well behaved time-series; naturally, all lags of volatility used in the regressions are also log-transformed. 3.2.2 Algorithmic trading The amount of algorithmic trading is measured as the percent of the overall trading volume that includes an algorithmic trader as either a maker or a taker; that is, the percent of trading volume where a computer was involved in at least one side of the trade. In addition, we also considered an alternative measure that separates the trading volume into our four di erent types of trades, and calculates the percent of total volume that each type represents. The four di erent types, as before, are the trades where both maker and taker are human, where the maker is a human and the taker is a computer, where the maker is a computer and the taker is a human, and where both maker and taker are computers. However, using this ner measure of algorithmic trading added little to the empirical results found for the simpler measure and we do not report those results. 3.2.3 Other control variables The additional control variables included in the regressions, represented by x i;t in equation (1), are discussed below. First, lagged values of the dependent variables are included to control for serial correlation. Realized volatility has a strong serial correlation even for distant lags (e.g. Andersen, Bollerslev, Diebold, and Labys, 23 and Bollerslev and Wright, 2). We follow the work of Andersen et al. (27). In particular, to control in a parsimonious manner for the serial correlation in volatility, which tend to stretch back many lags, we include the rst daily lag of volatility, the weekly lag of volatility, calculated simply as the average over the past ve business days and the monthly lag of volatility, calculated as the average over the past 22 business days. As argued by Andersen et al. (27), such a lag structure will capture most of the long-memory features of (logged) realized volatility, without imposing a vast number of parameters to estimate. 1 Second, as described in the context of the instrumental variable approach outlined above, the lagged values of realized volatility for the other currency pairs are also included as regressors; the same number of lags as for the own dependent variable are used in all regressions. Finally, in order to control for the large secular trends in the fraction of algorithmic trading, monthly time dummies are included in the regressions. 1 Alternatively, one could also model the realized volatility as a long-memory or fractionally integrated process. In this case, the long-memory parameter (d) is estimated and the fractionally di erenced realized volatility series is used in the analysis. The results from such a speci cation are qualitatively identical to those shown in the paper, and are not presented. 1
3.3 Empirical results The empirical regressions results are presented in Table 4. We present OLS results, the LIML-IV results using all ve currencies, and the LIML-IV results using only the three main currency pairs. Each speci cation is estimated with or without time dummies as outlined previously. The lag structure described above is included in all regressions. We report results for the sample starting in January 26 and ending in December 27. In order to save space, only the estimates of the coe cient in front of the fraction of algorithmic trading are presented. As described before, the speci cation with time dummies represents the most interesting and relevant one, whereas the one without time dummies is primarily included for completeness. In addition to the coe cient on the fraction of algorithmic trading, the results for the J test of overidentifying restrictions, which provides a test of the instrument validity, and the Stock and Yogo (25) F test of weak instruments, which tests instrument relevance, are reported for the IV regressions. Failure to reject with the J test provides some evidence of the validity of the instruments. The Stock and Yogo (25) F statistic, which is equivalent to the F statistic for the excluded instruments in the rst stage regression, tests whether the instruments are weak. Rejection of the null of weak instruments indicates that standard inference on the IV-estimated coe cients can be performed, whereas a failure to reject indicates possible size distortions in the tests of the LIML coe cients. The critical values of Stock and Yogo (25) are designed such that they indicate a maximal actual size for a nominal sized ve percent test on the coe cient. Thus, in the case with all ve currencies used in the IV estimation, a value greater than 5:44 for this F statistic indicates that the maximal size of a 5 percent test will be no greater than 1 percent, which might be deemed acceptable; the corresponding critical value in the three currency speci cation is 8:68. In general, the larger the F statistic, the stronger the instruments. The OLS results show that there appears to be a positive association, or correlation, between the level of volatility and the fraction of algorithmic trading in the market, with highly signi cant estimates in all but one currency. The OLS estimates are not likely to provide an unbiased estimate of the causal relationship and turning to the IV results, most signs of a relationship disappears. The F statistics for the IV estimation raise some warning signs, however. In the speci cation with time dummies, which is of primary interest, the null of weak instruments can typically not be rejected at a level that insures no more than a maximal size of 1 percent in the tests of the IV coe cient. Only for the euro-dollar currency pair are there no substantial signs of weak instruments and in this case the coe cient on algorithmic trading is insigni cant; the coe cient on algorithmic trading is insigni cant for all other currencies as well, when time dummies are included. 11
4 The price impact of algorithmic trading In the previous section, we investigated whether the presence of algorithmic trading increases exchange rate volatility. However the inference is complicated by the secular trend in both algorithmic trading and realized volatility, as well as by endogeneity complications. In this section we indirectly determine whether computer trades cause excess volatility and high-frequency noise in the exchange rate. To this end we estimate returnorder ow dynamics in a vector autoregressive (VAR) framework in the tradition of Hasbrouck (1991a). This procedure allows us to identify whose trades, computer or human, have a permanent impact on prices and to determine whether exchange rate prices are more likely to over-react to computer or human trades. We will interpret the price s over-reaction to a particular type of order ow as this particular type of order ow causing excess volatility. This interpretation is consistent with information-based models (dynamic learning models with informed and uniformed investors), where liquidity traders do not contribute to the price discovery process (do not have a permanent impact on prices) and prices temporarily over-react to this type of trades (e.g., Albuquerque and Miao (28)), thus create excess volatility in prices. In contrast, in these models, the under-reaction of asset price s to order ow is a natural consequence of the learning process. 11 4.1 VAR estimation Similar to Hasbrouck (1991a), we allow returns to be contemporaneously a ected by order ow, but there is no contemporaneous e ect of returns on order ow. We also allow U.S. macroeconomic news surprises to a ect both returns and order ow (Evans and Lyons, 28). In particular, we estimate the following system of equations for each currency i r it = r + JX r ijr it j + j=1 OF (l) it = OF l + JX j=1 LX OF ijl r it j + JX l=1 j= LX r ijlof (l) it JX OF ijl OF (l) it l=1 j=1 K j + X r iks kt + " r it; (2) k=1 j + K X k=1 OF ikl S kt + " (l) OF it : where L = 2 or 4 depending on the decomposition of the order ow; that is, OF (l) it represents the l th component of order ow in currency i at time t, where the order ow components are speci ed in the decompositions below. r it is the 1-minute exchange rate return for currency i at time t; OF it is the currency i order ow 11 We note that the over- and under-reaction of prices to a particular type of order ow is di erent from the over- and underreaction of prices to public news, which are both considered a sign of market ine ciency. In particular, order ow types are not public knowledge, so that agents cannot condition on these variables. 12
at time t decomposed in three di erent ways: de ned in the most disaggregate way, fhh; HC; CC; CHg, de ned according to who initiates the trade, fhh + CH; CC + HCg, and de ned according to whether there is any computer participation in the market, fhh; CC + HC + CHg; S kt is the macroeconomic news announcement surprise for announcement k de ned as the di erence between the announcement realization and its corresponding market expectation. We use the International Money Market Services (MMS) Inc. real-time data on the expectations and realizations of K = 28 U.S. macroeconomic fundamentals to calculate S kt. The 28 announcements we consider are listed in Table 5 and are similar to those in Andersen et al. (23, 27) and Pasquariello and Vega (27). 12 For a detailed description of the data we refer the reader to Andersen et al. (23). Since units of measurement vary across macroeconomic variables, we standardize the resulting surprises by dividing each of them by their sample standard deviation. Economic theory suggests that we should also include foreign macroeconomic news announcements in equation (2). However, previous studies nd that exchange rates do not respond much to non-u.s. macroeconomic announcements, even at high frequencies, e.g. Andersen et al. (23), so we expect the ommitted variable bias in our speci cation to be small. Following the tradition of the VAR price-impact literature we focus on the highest sample frequencies and estimate the VARs using the minute-by-minute data. The estimation period is restricted to the 26 27 sample, and the total number of observations for each currency pair is 717; 12 in the full sample and 89; 28 in the three month sub-sample (September, October and November of 27). In both samples, 2 lags are included in the estimated VARs, i.e. J = 2. Before considering the impulse response functions and the variance decomposition, it is worth brie y summarizing the main lessons from the estimated coe cients in the VAR. Since there are many coe cients estimated for each currency pair, we only report in Table 5 the macroeconomic news announcement coe cients and the contemporaneous order ow coe cients in the exchange-rate equation when we consider the most disaggregate decomposition of order ow: HH, CH, HC, and CC. The rest of the coe cient estimates are not shown but we brie y summarize our results below and we also report the impulse response function results. In addition to the coe cient estimates we report the R 2 of estimating the structural VAR with OLS equation by equation and the R 2 when we only consider news announcement times and run an OLS regression of 1- minute exchange rate returns on macroeconomic news announcements. This latter R-squared indicates how much do U.S. macroeconomic news announcements a ect exchange rate returns. As theory would predict, we nd that U.S. macroeconomic news announcements a ect less the euro-swiss and the euro-yen than the euro-dollar, dollar-yen and dollar-swiss franc exchange rates. 12 Our list of U.S. macroeconomic news announcements is the same as the list of announcements in Andersen et al. (27) and Pasquariello and Vega (27) with the addition of three announcements: unemployment report, core PPI and core CPI. 13
The rst own lag in all the order ow equations is always highly signi cant, and typically around :1 for all currency pairs. The main exception is the coe cient on the own lag in the computer-maker/computer-taker order ow regression, where the rst order autoregressive coe cient is typically much smaller and in the range :1 to :5. There is thus a sizeable rst-order autocorrelation in most of the order ow components, but less so in the computer-maker/computer-taker order ow; the higher order lags are generally substantially smaller, but typically positive. The coe cients on the rst order cross-lags in the order ow regressions are most often substantially smaller than the coe cient on the own lag and vary in signs. Lagged returns have a small but positive impact on human-maker/human-taker order ow, suggestive of a form of trend chasing in the order placement. Interestingly, the opposite is true for the computer-maker/computer-taker order ow, where the rst lag of returns always has a negative coe cient; for the other two order ows, the results are mixed across currencies. Finally, the return equation shows that minute-by-minute returns tend to be negatively serially correlated, with the coe cient on the rst own lag varying between :5 and :15; there is thus some evidence of mean reversion in the exchange rates at these high frequencies, which is a well-know empirical nding. All four order ows are signi cant predictors of returns. The price impact of the lagged order ows range from around 1 to 15 basis points per billion units of order ow (denominated in the base currency), as compared to a range of approximately 2 1 basis points in the contemporaneous order ow. The main di erences in the coe cients on the lagged order ows in the returns equation are between currencies rather than between the di erent types of order ows. predictor of returns than the others. That is, there is little evidence that one type of order ow is a better Again, the rst order lags dominate the relationship. It should also be stressed that despite the strongly signi cant estimates that are recorded in the VAR estimations and the relatively high R 2 reported in Table 5, the amount of variation in the order ow and return variables that is captured by their lagged values is very limited. The R 2 for the estimated equations with only lagged variables are typically around three to four percent for the order ow equations, and often less than one percent for the return equations. Again, the main exception is the computer-maker/computertaker order ow equation, which typically yields R 2 s of less than one percent. Overall, from examining the coe cients in the estimated VARs, there is little evidence that there is any systematic di erence between the di erent types of order ows in the way that they a ect the dynamics of returns. The most notable nding is probably the substantially lower persistence and predictability that is found for the computer-maker/computer-taker order ow. 14
4.2 Impulse Response Function Results In Table 6 we show a summary of the results from the impulse response analysis based on the full sample for 26-27, when the size of the shock is the same across the di erent types of order ow: one billion base currency shock to order ow. Because the standard deviation of order ow is di erent across maker/taker order ow pairs (shown in Tables 1 and 2) we also consider a shock that varies across the di erent maker/taker order ow pairs according to the average size shock. To that end we show in Table 7 the (cumulative) impulse response of returns to one standard deviation shock to a particular type of order ow. All the responses are measured in basis points. Each table has three panels in which we show the results from the three di erent order ow decompositions: de ned in the most disaggregate way, fhh; HC; CC; CHg, de ned according to who initiates the trade, fhh + CH; CC + HCg, and de ned according to whether there is any computer participation in the market, fhh; CC + HC + CHg. We show the short-run (instantaneous) impulse responses as well as the long-run cumulative responses, along with the long-run variance decomposition (Table 1 and 11). The long-run statistics are all calculated after 3 minutes, at which point the cumulative impulse responses have converged and can thus be interpreted as the long-run total impact of the shock. Figures 4 and 5 trace out the full paths of the cumulative impulse responses based on the most disaggregate decomposition of order ow, again using two di erent shock sizes: one billion base currency order ow shock and one standard deviation shock, respectively. Given the large sample sizes being used, there is little gain from showing con dence bands for the impulse response functions, as the coe cients are very precisely estimated. In general, the standard errors of the estimates are small enough that they contribute little to the analysis, and are not displayed in the tables either. Starting with a hypothetical shock of one billion base currency order ow, the results in Table 6 and Figure 4 show that, in general, HH order ow impacts prices more than CC order ow. However the di erences in price impact, although statistically signi cant, are not economically signi cant. In particular, the di erence in the responses across order ow types in the two currencies with the largest trading volume, the euro-dollar and dollar-yen markets, is very small, it ranges from 1 to basis points. Furthermore, there are some notable exceptions to this pattern. For instance the immediate response of the dollar-yen exchange rate to a CC order ow shock is almost 5 percent larger than the response to a HH shock of the same size. For the euro-swiss franc, the opposite is true. A similar picture emerges when we decompose order ow according to who initiated the trade, human-taker compared to computer-taker, and according to whether there is any computer participation, human-maker/human taker compared to computer-participation. In contrast to these results, the response to a hypothetical one standard deviation shock to the di erent order ows consistently shows that humans have a bigger impact on prices than computers (Table 7 and 15
Figure 5) and the di erences are economically signi cant. In particular, one standard deviation shock to HH order ow has an average long-run e ect of.5 basis points across currencies compared to one standard deviation shock to CC order ow which has an average e ect of.1 basis points. Similarly, we obtain that human-taker trades a ect prices on average by.6 basis points, while computer-taker trades a ect prices on average by.3 basis points. Interestingly, focusing in the disaggregate order ow decomposition, we observe that when humans trade with other humans they in uence prices more than when they trade with computers, and when computers trade with other computers they in uence prices less than when they trade with humans. Our interpretation is that computers provide liquidity more strategically than humans, so that the counterparty cannot a ect prices as much. We also nd that the price response to order ow varies across currencies as these markets di er along several dimensions. Trading volume is largest in the euro-dollar and dollar-yen markets, compared to the euro-yen market, and price discovery clearly happens mostly in the two largest markets. In the cross-rate market, euro-yen, computers have a speed advantage over humans in pro ting from triangular arbitrage opportunities, where prices set in the euro-dollar and dollar-yen markets are very brie y out of line with the euro-yen rate. Consistent with this speed advantage we observe that human-maker/computer-taker order ow has a larger price impact in the cross-rate market than in the other two markets Trading volumes in the dollar-swiss franc and euro-swiss franc markets tend to be close, with dollar-swiss franc volume a bit higher on average, but it is widely believed that price discovery occurs more often in the euro-swiss franc market than the dollar-swiss franc market. In this case HC order ow also has a slightly larger price impact in dollar-swiss franc than in euro-swiss franc. The dynamics of the VAR system help reveal an interesting nding aside from the level of the price impact of order ow: There is a consistent and often large short-run over-reaction to CC shocks. That is, as seen both in Tables 6 through 9 and Figures 4 through 7, the short run response to a CC order ow shock is always larger than the long-run response, and sometimes substantially so. The euro-dollar in the sample covering September, October, and November of 27 provides an extreme case where the initial reaction to a one billion dollar shock is a 21 basis point move, but the long-run cumulative reaction is just 4 basis points (Table 8). Interestingly, the opposite pattern is true for the HH order ow shocks, where there is almost always an initial under-reaction in returns. To the extent that exchange rates follow random walks over medium term horizons, these impulse response patterns thus suggest that CC trading might contribute to excess short-run noise or volatility. This over-reaction disappears when we consider only human-taker and computer-taker order ow, but it is still signi cant in the euro-dollar and dollar-yen market when we consider the HH and computer-participation (at least one computer counterparty) order ow decomposition. One possible interpretation could be that the participation of computers in these markets, in whatever form, 16
generates some excess short-run volatility. In addition to the impulse response functions, we also report the long-run forecast variance decomposition of returns in Table 1 and 11 for the full sample and the three-month sub-sample, respectively. 13 That is, within the framework of the VAR, what fraction of the total (long-run) variance in returns can be attributed to innovations in the di erent order ows. As originally suggested by Hasbrouck (1991b), this variance decomposition can be interpreted as a summary measure of the informativeness of trades, and thus, in the current context, a comparison of the relative informativeness of the di erent types of order ow. Consistent with the impulse response functions to one standard deviation shock to order ow, there are obvious patterns in the variance decompositions. The HH order ow makes up the dominant part of the variance share in most cases, which is not surprising given that this component constitutes the largest share on average across the sample period. In the last three months of the sample, this share has generally decreased. The share of variance in returns that can be attributed to the CC order ow is surprisingly small, especially in the latter sub-sample, given that this category of trades represent a sizeable fraction of overall volume of trade during the last months of 27, as seen in Figure 3. The mixed order ow (CH and HC order ow) typically contribute with about the same share to the explained variance. Overall, about 15 to 35 percent of the total variation in returns can be attributed to shocks to the four order ows. However, in most currency pairs, very little of this ultimate long-run price discovery that occurs via order ow does so via the CC order ow. The seemingly disproportionately small fraction of the explained return variance that can be attributed to the CC order ow is likely a result both of the generally smaller responses by returns to shocks from this order ow component, as seen in the impulse response analysis, as well as the generally smaller shocks that occur in this order ow as seen from the estimates of the standard deviation in the di erent order ows, 14 ;15 presented in Table 1. However, the (cumulative) impulse response functions to one-billion base currency shocks suggest that it is more due to the latter than to the former. 4.3 Summary Our empirical analysis provides three important insights. First, we nd that human order ow accounts for most of the (long-run) variance in exchange rate returns, i.e., humans are still the informed traders 13 The variance decompositions are virtually identical in the short- and long-run and thus we only show the long-run decomposition results. 14 The variance decompostion is a function of the (squared) terms in the Vector Moving Average (VMA) representation of the VAR and the variance of the shocks in the VAR equations (i.e. the variance of the VAR residuals). For a given shock size, the impulse response functions are a function of the (non-squared) VMA coe cients. 15 Strictly speaking, the variance decomposition is a function of the variance in the shocks in the VAR residuals and not in the original data entering the VAR, i.e. the variance of the unexpected shocks. However, since the R 2 s in the VAR equations are small, the variance in the VAR residuals and the original data are very similar. 17
in these markets. This can probably be attributed in part to the fact that investors are more likely to use algorithmic trading, relative to human trading, for the optimal execution of large orders at a minimum cost. Algorithmic trades appear to be successful at that task, so that computers break up the orders so as to have a small impact on prices. Second, we nd that, on average, computer-takers or human-takers that trade with a computer-maker do not impact prices as much as they do when they trade with a human-maker. result is that computers place limit orders more strategically than humans do. One interpretation of this This nding dovetails with the literature on limit order books that relaxes the common assumption that liquidity providers are passive. 16 Third, we show that there is an initial under-reaction to human-maker/human-taker order ow, while there is an initial over-reaction to computer-maker/computer-taker order ow. To the extent that there is an initial over-reaction to computer-maker/computer-taker order ow, we conclude that algorithmic trading may lead to some excess short-run volatility. There is also some evidence of over-reaction to order ow when computers participate in the market either as liquidity providers or liquidity demanders, but only for the euro-dollar and the dollar-yen markets. 5 How Correlated Are Algorithmic Trades and Strategies? We investigate the proposition that AT agents tend to have trading strategies that are more correlated than those of human agents. Since the outset of the nancial turmoil in the summer of 27, multiple articles in the nancial press have suggested that AT programs tend to be similarly designed, leading them to take the same side of the market in times of high volatility, and potentially exaggerating market movements. If AT (computer) agents and human agents trade randomly, then we should expect to see them trading with each other in proportion to their relative presence in the market. If, on the other hand, computer agents tend to have more homogeneous trading strategies, we should expect to see them trading less among themselves and more with human agents. At the extreme, if all computer agents used the very same algorithms and had the exact same speed of execution, we would expect to see no trading volume among computers. Therefore, the fraction of trades conducted between computers agents contains information on how correlated their strategies are. To test this question, we assume a simple market model in which computer agents and human agents trade randomly, and then compare the implications of that model to the actual data. In this model, there are two separate types of agents: makers and takers. Within each of these groups, 16 For example, Chakravarty and Holden (1995), Kumar and Seppi (1994), Kaniel and Liu (26), Goettler, Parlour and Rajan (27) allow informed investors to use both limit and market orders. Bloom eld, O Hara and Saar (25) argue that informed traders are natural liquidity providers and Angel (1994) and Harris (1998) show that informed investors can optimally use limit orders when private information is su ciently persistent. 18
there are both computer agents and human agents. During any given period k, computer agents make up some xed proportion m;k of makers and some xed proportion t;k of takers. We allow these proportions to di er from one another and to vary between periods. The remaining makers and takers are human agents, in proportions (1 m;k ) and (1 t;k ), respectively. The model abstracts from the fact that, in practice, actual traders can act as both makers and takers. At each time k, we allow the agents to trade. We assume that any order submitted by a taker will be randomly matched with a maker, regardless of the identity of either party as a computer or a human. Thus, trading among computer agents, among human agents, and between computer agents and human agents should occur in proportion to those agents relative presence on either side of the market. We use the following nomenclature to refer to the four types of trading volume during period k: V olhh k refers to trading volume between human makers and human takers, V olhc k to trading volume between human makers and computer takers, V olch k to trading volume between computer makers and human takers, and V olcc k to trading volume between computer makers and computer takers. Similarly, P cthh k, P cthc k, P ctch k and P ctcc k refer to those volumes expressed as a percent of total trading volume. Under our random trading model, the following conditions hold: P cthh k = (1 m;k ) (1 t;k ) P cthc k = (1 m;k ) t;k P ctch k = m;k (1 t;k ) P ctcc k = m;k t;k Since m;k and t;k may vary over time, the model provides no guidance as to the expected levels of P cthh k, P cthc k, P ctch k and P ctcc k. But using the assumption of random trading between AT agents and human agents, the model does provide some guidance as to the proportions of these quantities. In particular, the model implies that V olhh k V olch k = V olhc k V olcc k ; for all strictly positive values of m;k and t;k. For simplicity of notation, we de ne two key measures: RH k = V olhh k V olch k, the human taker ratio RC k = V olhc k V olcc k, the computer taker ratio Intuitively, for each agent type, the taker ratio is an expression of the propensity of takers of that type to trade with human makers relative to computer makers. In a market with mostly human makers, we would expect these ratios to exceed 1, while in a market with mostly computer makers, we would expect these ratios to be smaller than 1. 17 17 The use of this model, and the appeal to these ratios, is not an attempt to complicate a simple analysis. We only know the volume of completed trades in each category (HH, CC, HC, CH) ex-post, but not the number of trades attempted by each type of trader. This limits the number of variables relative to the number of unknowns we seek to nd, which allows us to make 19